Unverified50% confidenceFactExact time

Blackwell将FP4引入推理，配合微缩放技术可将权重存储压缩至FP16的1/4

1

Sources

50%

Confidence

Medium-term (~90 days)

Relevance

7/15/2026

First Seen

Valid until: 10/13/2026

Sources

软硬协同设计：构建硬件友好的LLM架构指南

Related Claims

UnverifiedINT4 相较 FP16 实现了 4 倍存储压缩75% similar Verified4-bit量化使模型存储空间缩减为FP16的约1/475% similar UnverifiedFP4 相对 FP16 可节省75%存储，采用每组16或32个权重共享一个 FP8 缩放因子的分组量化策略74% similar Unverified量化KV缓存（如FP8或INT4存储KV向量）可将缓存大小压缩到原来的1/2甚至1/472% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/512134

API

curl https://kongchang.com/api/v1/knowledge/claims/512134

MCP

get_claim(id=512134)